陈志航·ZH
Family中

AI for Drug Discovery · Computational Biology

Machine learning for drug discovery and molecular design

Ph.D. candidate in Biology at Tsinghua Shenzhen International Graduate School, jointly trained at the Shenzhen Medical Academy of Research and Translation under Prof. Yaoqi Zhou; M.Sc. in Medicinal Chemistry from Fudan University, supervised by Renxiao Wang and Yifei Qi. I develop computational methods for protein and peptide design, protein/RNA language models, and nucleic acid drug evaluation. My industry collaboration includes independent responsibility for peptide sequence design for the CGB3002 blood–brain barrier program, which has since entered clinical development.

Explore researchGet in touch
AIProtein design
RNAStructure & regulation
MLScientific machine learning

Education

Academic training

Tsinghua Shenzhen International Graduate School

Ph.D. candidate in Biology

School of Pharmacy, Fudan University

M.Sc. in Medicinal Chemistry

School of Pharmacy, China Pharmaceutical University

B.Sc. in Pharmacy

Publications & manuscripts

Publication details →
08
2026ManuscriptUnder review

Cooperative Active-Site Remodeling Couples Catalytic Efficiency and Stereoselectivity in L-Threonine Aldolase

Co-first author

Performed quantum-mechanical (QM) calculations, molecular dynamics (MD) simulations, and trajectory analysis to investigate the structural and energetic basis of active-site remodeling in L-threonine aldolase, integrating local electronic-energy comparisons in enzyme-cluster models with analyses of complex stability and conformational fluctuations.

Projects & research experience

Research projects →

CGB3002: peptide design for blood–brain barrier delivery

Independently designed peptide sequences and assessed their structures using multiple deep-learning approaches. CGB3002 has since entered Phase I clinical development for Alzheimer's disease; my contribution was to the earlier computational design stage. Sequences and unpublished design details remain confidential.

ChainGen Biopharma · Independently led computational design

AI-assisted peptide substrate design for rapid SERS sensing

Led machine-learning-assisted sequence and structure design of MMP-2-cleavable peptide substrates to support matched dual-enzyme kinetics and rapid detection. The designed substrates were incorporated into the team’s dual-enzyme SERS sensing array, evaluated for intraoperative chemical mapping of epileptic foci, linking computational molecular design with diagnostic research.

Co-first author (second-listed) | Computational substrate design

Binding-affinity prediction for cytokine receptor interfaces

Used deep learning to predict binding free-energy effects across single-residue saturation substitutions at a cytokine receptor complex interface involving the IL-2 receptor alpha chain.

Independent computational project

Computational design of peptide substrates for palmitoyltransferase

Applied deep-learning methods to peptide sequence redesign and computational substrate design.

Independent computational project

Peptide binder design around an ATP1A3 phosphorylation site

Conducted computational peptide binder design focused on the S643 phosphorylation site of the sodium/potassium-transporting ATPase alpha-3 subunit (ATP1A3).

Independently led computational design

SSDE: RNA candidate discovery through iterative learning

Developing a discovery workflow that identifies RNA structural candidates from large collections of natural sequences when experimentally resolved structures are scarce. Responsible for data curation, model development and training, computational evaluation, candidate analysis, and delivery. The project has progressed to experimental validation.

Self-Sustained Discovery Engine · End-to-end computational lead

Reliability and cross-source generalization of RNA design models

Built a framework integrating multiple models and structural evidence for functional RNA evaluation, with data provenance checks, cross-source comparisons, and retrospective evaluation under a fixed budget. Interim findings show that ensemble gains depend on the data source, additional models do not consistently improve ranking, and structural plausibility cannot substitute for evidence of function.

Evaluation framework · Interim findings established

Antisense oligonucleotide activity prediction and benchmarking

Curated approximately 150,000 public records and benchmarked sequence and chemical features against RNA language-model representations. Used evaluation splits that separate targets, data sources, and similar sequences. Completed multi-seed benchmarks and ablation studies to characterize calibration shifts and limits to generalization under distribution shift.

Machine learning for nucleic acid therapeutics · Initial benchmarks completed